Fig. 5: Sliding window, real-world progression analysis, and retrospective analysis.

The heat maps show different permutations of training and test sets to assess model robustness or degradation over time. Performance is evaluated using Area Under the Receiver Operating Characteristic curve (AUROC). The color-coding shows the model performance for each training-validation pair. For the sliding window experiment (a), the model is trained on a moving three-year data span (vertical axis) and then evaluated as the temporal gap between training and validation years widens (horizontal axis, longevity analysis). The progression analysis (b) simulates a scenario where the model is implemented from the outset, with annual retraining using all available data up to that point, and prediction on a single future year. This approach mimics the real-world process of continuously updating the model as new data emerges each year. Specifically, the triangle heatmaps in a and b offer three readings: first, the values on the diagonal display model performance on the year directly following the training period; second, moving along the horizontal axis highlights changes in model performance as the temporal gap between training and test years widens; third, moving along the vertical axis depicts how model performance is influenced when training the model on increasingly more recent years, but testing on the same data from a single future year. The final heatmap (c), shows the incremental learning in the reverse direction, i.e., when adding a fixed set of 1000 training samples from each previous year.